Handling the Positive-Definite Constraint in the Bayesian Learning Rule